A full repro and instructions available at https://github.com/guptadeepanshu/r8Issue

Steps to reproduce:

Build mylib (a library with obfuscation enabled and AGP 3.5.3 or older):

cd mylib
./gradlew publishToMavenLocal

Build and install the release variant of the app (uses the mylib from above, and uses AGP 3.6.0 or above):

cd myapp
./gradlew installRelease

On the device, enter any text that contains "FIRST", "SECOND" or "THIRD".
Press "Submit" button

Expectation:

If the text contains one of the three keywords, it should display a message like this is the first enum value.

Observation:

No match found is shown for any text that contains only words in caps, like FIRST. If the text contains b, the message corresponding to first is shown. Similarly, c shows second and d shows third.

More details:

The issue started due to https://r8-review.googlesource.com/c/r8/+/40221. The first AGP to include it was 3.6.0. The issue exists in all versions since then.

The mylib is compiled with 3.5.3. It is also obfuscated before distribution. Since the above change was not present in 3.5, the enum MyEnum is not inlined in Api class. However, the enum itself is obfuscated. This result is a strange class where the decompiler shows the enum to be of the below form:

package com.example.mylib;

public enum a {
    b("this is the first enum value"),
    c("this is the second enum value"),
    d("this is the third enum value");

    public final String a;

    public a(String var3) {
        this.a = var3;
    }
}

However, at runtime, calling a.b.name() returns FIRST and not b. I'm not an expert in the bytecode specifications. So, I don't know exactly how it's happening.

When such an enum is used in the compilation of the app, where enum name inlining is enabled, it ends up inlining a call to name() with the obfuscated name. So, a.b.name() becomes b instead of FIRST.

So, there are two workarounds that are available for the app right now:

Use AGP 3.5.3 or older
Disable R8 and use proguard (set android.enableR8=false in gradle.properties)

It would be greatly appreciated if the inlining was able to handle this obfuscated enums produced by older versions of R8.

Issue 160351050

Links (20)

Issue metadata